This paper presents a novel and effective technique based on Artificial Neural Net (ANN) technology to reverse map from process response to the input (control) variables when the input-response relationships are nonlinear, complex or intractable by theory. This is often the problem when the control variables in manufacturing or prototype development must be set such that the response hits a specified target. The ANN technique avoids countless empirical searches in the decision space and thus minimizes the expenditure on R&D or production resources. Additionally, this paper illustrates how one may build an ANN model with top-flight performance in Microsoft Excel® when a commercial neural net software is unavailable. The paper includes a learningoriented reworking of a well-established example from the response surface literature. In conclusion, it indicates room for further research on training data collection methods.
Keywords: Process modelling, Artificial neural networks, Connection weights, Optimization, Reverse mapping, Regression
1. Process Optimization using Empirical methods
This paper could be retitled, for it asks “I have a process response (‘the answer’) that is now sitting at its desired target value. But how did the engineer reach it? What settings of the design or process control variables did she use (‘the question’) to deliver the response on that target?” One must note that this is often a critical question in engineering design and process control.